wchevreuil commented on code in PR #8221:
URL: https://github.com/apache/hbase/pull/8221#discussion_r3248930878
##########
hbase-balancer/src/main/java/org/apache/hadoop/hbase/master/balancer/CacheAwareLoadBalancer.java:
##########
@@ -88,10 +90,11 @@ public class CacheAwareLoadBalancer extends
StochasticLoadBalancer {
private Long sleepTime;
private Configuration configuration;
- public enum GeneratorFunctionType {
- LOAD,
- CACHE_RATIO
- }
+ private float lowCacheRatioThreshold;
+ private float potentialCacheRatioAfterMove;
+ private float minFreeCacheSpaceFactor;
+
+ private BigDecimal simulatedRatio = new BigDecimal(0);
Review Comment:
It's not thread safe for sure, but as of now, balancer is a single thread
only operation (no multiple parallel balancers run at the same time).
Unfortunately, we need to keep this global variable in order to rollback the
cache ratio increment if the plan isn't accepted.
In StochasticLoadBalancer.balanceTable we do:
1) Pick a generator and cost function at random;
2) Call generate() on the generator so that we set that variable to 0 and
get a move plan;
3) Simulate the plan, which calls regionMoved(). If the chosen cost function
was the CacheAwareCostFunction, we will calculate the impact of that plan on
the overall cache ratio. This impact is what we save simulatedRatio field;
4) Calculate the total cost with this plan. If it's lower than the lowest
cost so far, keep it. Otherwise, roll it back by using the same cost function
but inverting the to/from servers order. If it's the CacheAwareCostFunction,
regionMoved would be called and it would enter the branch that reverts the
value from the cacheRatio.
5) Iterate again. Reset the variable on step 2;
Without this, we would need to always recalculate all individual
regions/servers ratios on every new plan simulation (every iteration of the
StochasticLoadBalancer loop), which is costly.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
